Tropical cyclone tracks and environmental fields from a 1,000-member deep-learning climate emulator ensemble

Dataset Description This dataset supports the analysis presented in the paper titled "On the seasonal predictability of the 2020 North Atlantic tropical cyclone season" by Levin et al. The preprint is available at: https://doi.org/10.31223/X5CN1R The dataset contains tropical cyclone (TC) track data and monthly mean large-scale environmental fields derived from a 1,000-member ensemble generated using the ACE2 deep-learning climate emulator. The ensemble covers the years 1982 and 2005–2020, with each member representing a physically plausible realization of atmospheric variability under observed sea surface temperature (SST) forcing for a given year. IMPORTANT DATA AVAILABILITY NOTE Due to storage constraints on Zenodo, this repository contains only the subset of data corresponding to the 2020 TC season. The full dataset, including all other years (1982 and 2005–2019), is hosted externally and can be accessed at: https://tigress-web.princeton.edu/~GEOCLIM/el2358/ If you use this dataset, please cite the Zenodo DOI associated with this record (10.5281/zenodo.19456834), including when using data accessed from the external repository. 1. TC Track Data TC tracks are provided as .txt files and were identified using the TempestExtremes tracking algorithm. Each track includes 6-hourly information for global TCs: · Storm location (latitude and longitude) · Minimum sea level pressure (hPa) · Maximum 10m wind speed (m/s) 2. Monthly Mean Environmental Variables The dataset also includes global monthly mean environmental fields at 1° × 1° resolution, stored in NetCDF format. Variables include: · eastward_wind_6 : Zonal wind component at sigma pressure level 6 (~800 hPa) in m/s · eastward_wind_3: Zonal wind component at signal pressure level 3 (~250 hPa) in m/s · northward_wind_6: Meridional wind component at sigma pressure level 6 in m/s · northward_wind_3: Meridional wind component at sigma pressure level 3 in m/s · PRATEsfc: Surface precipitation rate in kg/m2/s · surface_temperature: surface temperature in K · Specific_total_water_5: Specific total water (vapor + condensates) at sigma pressure level 5 (~600 hPa) in kg/kg Ensemble Generation Methodology For each target year, a 1,000-member ensemble was constructed using an autoregressive integration framework with fixed SST boundary conditions: · Ten independent simulations (“chunks”) were run in parallel, each initialized by a different initial condition. · Each chunk consists of a 100-year integration with identical, repeated observed SST forcing for the target year. · Each simulated year within a chunk is treated as an independent ensemble member. Thus, each year’s ensemble is distributed across 10 folders, each containing 100 realizations. Important note on time dimension: The model output uses a continuous time axis that extends beyond the target year (e.g., 2020–2120 for a 100-year simulation initialized in 2020). This extended timeline does not represent a forecast into the future. Instead, each simulated year should be interpreted as an independent realization of the same target year’s climate state under identical SST forcing. The forward integration is a technical feature of the model and enables efficient sampling of internal atmospheric variability. Code Availability The code to run one year’s 1,000-member ensemble and process the output, and track TCs is available at https://github.com/emmalevin/ACE2_1000